Cookbook recipes to get up and running with Spice.ai quickly 🚀
No Spice release includes the llama.cpp GGUF engine this recipe needs yet; until one does, follow the build steps in the base recipe (opens in a new tab).
This directory is the GLM recipe (opens in a new tab) with the model swapped for huihui-ai/Huihui-GLM-4.7-Flash-abliterated (opens in a new tab), a version of GLM-4.7-Flash with its refusal behavior removed by abliteration, in the importance-matrix i1-Q4_K_M GGUF (18.1 GB) published by mradermacher (opens in a new tab). Everything else is unchanged.
Abliteration removes the "refusal direction" from a model's weights, so the model answers requests the base model would decline. huihui-ai describes its method as a proof-of-concept implementation of refusal removal (opens in a new tab). The result has none of the base model's safety training. Use it for research, red-teaming and controlled experiments, add your own moderation before exposing it to users, and review its outputs. huihui-ai's model card carries its own usage warnings and disclaims responsibility for use.
Wait for Model [glm-4.7-flash-abliterated] deployed, ready for inferencing, then in a second terminal:
The API serves the model as glm-4.7-flash-abliterated at http://localhost:8090/v1/chat/completions, with the same reasoning behavior as the base model. On an M3 Max with the trunk Metal build it loaded after a 4 min 44 s download, thought by default ("Reply with only the integer 1729." took 120 tokens with a 420-character trace, the same at low), answered in 2 tokens with reasoning_effort: none, and decoded at 77 tokens per second.
The mradermacher repository holds i1-IQ3_M (13.2 GB, 24 GB Macs), i1-Q4_K_M (18.1 GB, the default, 32 GB and up), i1-Q5_K_M (21.3 GB, 36 GB and up) and i1-Q6_K (24.6 GB, 48 GB and up), and the companion non-imatrix repository (opens in a new tab) adds Q8_0 (31.8 GB, 48 GB and up). These follow the GPU budget described in the base recipe's Memory (opens in a new tab) section; a 16 GB Mac's 10.7 GiB budget leaves no room for any of them.
Two abliterated builds of GLM-5.3-Flash exist. Both need the memory described in the base recipe's GLM-5.3-Flash section (opens in a new tab).
huihui-ai/Huihui-GLM-5.3-Flash-abliterated-GGUF (opens in a new tab) is Unsloth's quantization ladder with the ablated layers (15 to 35) rewritten in the first shard, so its UD-IQ2_XXS build is the same 101.8 GB as the base model's and fits a 128 GB Mac:
orcarouter/GLM-5.3-Flash-Uncensored-GGUF (opens in a new tab) documents its abliteration with measurements (harmful-prompt refusal from about 89% to the low teens, 0% over-refusal, and per-quant perplexity and top-1 agreement), and drops the NextN layer, so it decodes without drafting. Its smallest build, Q2_K, is 116.9 GB in three shards, which exceeds the 107 GiB GPU budget of a 128 GB Mac; it is a build for a 192 GB or larger Mac, or for a 128 GB Mac after raising the GPU memory limit with sudo sysctl iogpu.wired_limit_mb=<MB> (the setting resets on reboot, and leaves macOS with what remains). The repository is gated: sign in to Hugging Face, accept the model's license on its page, and add hf_token: ${ secrets:HF_TOKEN } to params, with a read token as HF_TOKEN in .env.
huihui-ai/Huihui-GLM-5.2-abliterated-GGUF (opens in a new tab) is the abliterated GLM-5.2 (753B). Its smallest build, UD-IQ1_S_MXFP4, is 216 GB in six shards and its UD-IQ1_M 231 GB, so like the base model it is for a 512 GB Mac Studio, not a MacBook Pro (a 256 GB Mac's GPU budget of about 192 GiB is below these files). The spicepod shape is the same: list the shards in files.